Global Navigation Satellite System (GNSS)-based navigation is usually considered as not usable indoors where the satellite visibility is degraded, and the complex propagation conditions perturb the GNSS signals with reflection and refraction. However, this paper presents a novel approach called LIGHT(Light Indoor GNSS macHine-learning-based Time difference carrier phase) that can select healthy indoor GNSS carrier phase data thanks to Machine Learning (ML) for positioning. The selected carrier phase data are fed into a Time Difference Carrier Phase (TDCP) based Extended Kalman Filter (EKF) to estimate the user’s velocity. Two indoor scenarios (shopping mall and railway station) are tested over a 2 km total walking distance. It is shown that at least half of the epochs become usable for GNSS TDCP standalone positioning, and the accuracy of the velocity estimates can improve up to 87% in terms of the 75th percentile of the absolute horizontal velocity error compared with the non-ML approach. Furthermore, a newly-designed hybridization filter LIGHT-PDR that fuses the LIGHT algorithm and Pedestrian Dead Reckoning (PDR) solution together is applied to perform seamless indoor/outdoor positioning in a more robust way.